24964db1b986850ea5a194bd7be2d30a1f256e50,slm_lab/agent/algorithm/actor_critic.py,ActorCritic,train_shared,#ActorCritic#,217

Before Change

            policy_loss = self.calc_policy_loss(batch, advs)  // from actor
            val_loss = self.calc_val_loss(batch, v_targets)  // from critic
            loss = policy_loss + val_loss
            self.net.training_step(loss=loss, lr_clock=self.body.env.clock)
            // reset
            self.to_train = 0
            self.body.entropies = []
            self.body.log_probs = []

After Change

        Trains the network when the actor and critic share parameters
        loss = self.policy_loss_coef * policy_loss + self.val_loss_coef * val_loss
        """
        clock = self.body.env.clock
        if self.to_train == 1:
            batch = self.sample()
            with torch.no_grad():
                advs, v_targets = self.calc_advs_v_targets(batch)
            policy_loss = self.calc_policy_loss(batch, advs)  // from actor
            val_loss = self.calc_val_loss(batch, v_targets)  // from critic
            loss = policy_loss + val_loss
            self.net.training_step(loss=loss, lr_clock=clock)
            // reset
            self.to_train = 0
            self.body.entropies = []
            self.body.log_probs = []

Instances